Capacity Allocation at the Source: Sparse Target Optimization for LLM Knowledge Editing
Authors: Dongyao Chen, Jiayun Lei, Zhiyuan Ji, Zhiying Deng, Wei Liu, Xiaobo Sun
Organizations: Renmin University of China · Laboratory for Artificial Intelligence and New Forms of Education, Central China Normal University · School of Computing, National University of Singapore · Department of Human Genetics, Emory University School of Medicine
Knowledge editing has emerged as a promising approach for efficiently updating embedded knowledge in large language models (LLMs). It first computes an ideal target hidden state that steers the LLM toward the new output, and then edits the model parameters so that the original hidden state is mapped to this target. While most existing methods focus on the second stage of parameter editing in preserving other knowledge, we find that the upstream construction of the target hidden state also plays an important role in allocating the model's limited capacity space. We first establish a theoretical connection between interference in the parameter space and interference between target hidden states, showing that the two are positively correlated. Building on this insight, we propose a simple yet effective ℓ1 regularization method that prunes away unnecessary ``tentacles'' of each knowledge vector, retaining only the essential components. With fewer extraneous tentacles, each edit affects less other knowledge, enabling better coexistence. Extensive experiments across multiple LLMs and benchmark datasets verify the effectiveness of our method. Code: https://github.com/wintersdragon-c/STO.
Figures & tables
Figure 1: Overview of locate-then-edit model editing. The method first finds a target hidden state for the decisive token at the decisive layer via gradient descent (we call it upstream target construction), then edits a key subset of parameters to steer the original hidden state ( m0 ) toward this target ( m ), producing the desired output ( Ytarget ) (we call it downstream parameter updating).
Table 1: The results on the MCF and ZsRE datasets with GPT2-XL.
Table 2: The results on the MCF and ZsRE datasets with Qwen2.5-7B-Instruct.
Table 3: Compatibility of STO with PRUNE, RECT, and AlphaEdit on GPT2-XL (2,000 edits per dataset). Each combination’s absolute results are followed by Δ (combination minus its baseline in Table 1 ), computed before rounding. Changes are in percentage points except for DKL ; bold Δ values indicate improvements.
Figure 2: GPT2-XL MCF scaling to 10000 knowledge items.
Figure 3: Verifying feature sparsification. (a) We use varying thresholds (x-axis) to define near-zero values; the y-axis shows the proportion of dimensions falling below each threshold. (b) Distribution of per-dimension values: the x-axis denotes the value of each dimension in the residual, and the y-axis denotes density.
Figure 4: Dimension-level ∣rw/oℓ1rsto∣ . We sort the dimensions of r in descending order by absolute value and compute ∣rw/oℓ1rsto∣ . For ease of visualization, we group every 100 adjacent dimensions and report the average ratio per group.
Method
∣ra⊤rb∣ (mean) ↓
∣ra⊤rb∣ (median) ↓
CE (mean)
CE (median)
STO
251.2152
195.9598
0.0161
0.0021
w/o ℓ1
541.1298
442.3562
0.0152
0.0020
Table 4: We report the mean/average cross-entropy (CE) loss and sample 500K (a,b) pairs to report ∣ra⊤rb∣ . We achieve much lower knowledge interference (Eq. 12 ) but only marginally higher CE.
Table 5: The results on the MCF and ZsRE datasets with GPT-J.
Table 6: Qwen2.5-7B-Instruct 10000-edit comparison. The MCF block reports completed runs on on the same exact 10,000-edit manifest.
Table 7: GPT2-XL 10000-edit comparison. The MCF block reports completed runs on on the same exact 10,000-edit manifest.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Table 8: Compatibility of STO with PRUNE, RECT, and AlphaEdit on Qwen2.5-7B-Instruct (2,000 edits per dataset). Each combination’s absolute results are followed by Δ (combination minus its baseline in Table 2 ), computed before rounding. Changes are in percentage points except for DKL ; bold Δ values indicate improvements.
Model
Dataset
λ1
GPT2-XL
MCF
9×10−4
GPT2-XL
ZsRE
6×10−4
Qwen2.5-7B-Instruct
MCF / ZsRE
3×10−4
GPT-J-6B
MCF / ZsRE
3×10−4
Appendix
Table 9: Selected sparsity coefficients for the standard MEMIT + STO experiments.
λ1
Eff-S ↑
Eff-A ↑
Gen-S ↑
Gen-A ↑
DKL↓
Top-1 ↑
GPT2-XL / MCF
0
92.3
75.0
80.0
43.3
0.85
62.9
3
98.6
95.8
92.2
60.0
0.97
63.0
6
97.8
93.6
89.2
53.7
0.85
65.2
9
95.3
87.6
84.3
46.3
0.76
67.1
GPT2-XL / ZsRE
Appendix
Table 10: Sensitivity to the sparsity coefficient with 2,000 edits per setting. Coefficients in the first column are in units of 10−4 ; bold coefficients mark the selected operating points.
Knowledge editing aims to efficiently update factual information in Large Language Models (LLMs) without full retraining. However, existing methods still suffer from performance degradation in batch knowledge editing. We identify that semantic representation entanglement, such as overlapping concepts and shared syntactic patterns, accumulates interference in the representation space and reduces editing precision. To bridge this gap, in this paper, we propose Orthogonal Representation Editing (ORE), which performs edits in the hidden representation space of LLMs by constructing a general semantic subspace and enforcing orthogonal constraints on edit vectors, effectively decoupling semantic entanglement. Furthermore, we introduce a gated non-linear representation head to enable adaptive learning of editing locations and precise control over knowledge injection. Extensive experiments show that ORE outperforms existing methods and achieves superior performance in cross-lingual knowledge editing scenarios. We release our code at https://github.com/YVVH/ORE.
Wenhao Yu, Zhicong Lu, Bo Lv +4
School of Computer Science and Technology, Tianjin University · Kexin Technology · University of Chinese Academy of Sciences +2
Parameter-based knowledge editing updates the internal knowledge of large language models (LLMs) via localized weight modifications and has attracted significant attention. However, most existing methods overlook fundamental theoretical limitations and are rarely evaluated under realistic, practice-oriented settings. In this paper, we first present a theoretical analysis based on the dimensional Collapse Hypothesis, explaining how localized parameter edits can propagate along fragile directions in the representation space, inducing global interference and ultimately causing reasoning collapse. Building on this insight, we conduct a comprehensive empirical evaluation by systematically varying knowledge complexity, number of edits, evaluation dimensions, and baseline methods. Our results show that parameter-based editing methods consistently damage core LLM capabilities. In contrast, a simple retrieval-based baseline achieves consistently stronger performance than all parameter-editing methods across all evaluated conditions. These findings highlight that preserving the fundamental capabilities of LLMs after knowledge editing should be a central concern for future research.
Wanying Ren, Xin Song, Futing Wang +2
East China Normal University, Shanghai, China · Westlake University, Hangzhou, China · Nanyang Technological University, Singapore
Sequential knowledge editing in Large Language Models (LLMs) often leads to a degradation of general capabilities due to cumulative parameter interference. In this paper, we revisit prominent projection-based methods like AlphaEdit and reveal a surprising mismatch: their robustness is primarily driven by update regularization rather than the intended null-space projection. Based on this insight, we propose FisherEdit, a theoretically grounded method built on Fisher geometry, which replaces coarse activation-space projections with importance-aware regularization in parameter space. It penalizes changes to sensitive parameters more strongly while retaining flexibility for new facts, and admits an efficient row-wise closed-form solution. Experiments on Llama-3, Qwen2.5, and Olmo-3 across multiple benchmarks demonstrate that FisherEdit achieves state-of-the-art performance while better preserving model stability and downstream task accuracy.
Yizhou Yang, Jiahong Fu, Zhiwu Wang +6
Xi’an Jiaotong University · Zhejiang University · City University of Hong Kong +2